A. Design of the dry EEG sensors
The proposed dry foam-based EEG sensor was specific designed to contact the skin of the forehead with the use of a conductive polymer foam made of a urethane material with a compression set of about 5~10%, as shown in Figures 1A and 1B. The conductive foam was covered with a 0.2-mm-thick taffeta material that was made from an electrically conductive polymer fabric (conductivity of about 0.07 ohm/cm2) and was coated with Ni/Cu on all of its surfaces to establish an electrical contact that was similar to that of silver EEG sensors. A 0.2-mm layer of Cu was used as an adhesion layer that was then connected to the wireless EEG acquisition module. The proposed dry foam EEG sensors were 20 × 20 × 9 mm3. The corresponding design specifications and the equivalent circuit of the skin-sensor interface for dry EEG sensors can be found in our previous study [35].
B. Wireless EEG acquisition module
Figure 1C and Figure 2 show the wireless EEG acquisition module and its application to gaming control. It was used to acquire EEG signals from the dry EEG sensors and included the (INA2126, Texas Instruments, USA), an acquisition component (AD8609, Analog Devices, USA), a microprocessor component (MSP430, Texas Instruments, USA), and a wireless transmission component (BM0403, Unigrand Ltd., Taiwan) [36]. To amplify and filter the EEG signals, a pre-amplifier, a band-pass filter (0.5~50 Hz) and an analog-to-digital converter (ADC) were embedded into our circuit board as a bio-signal amplifier and acquisition component modules. The gain of the amplifier and acquisition component was set to approximately 5500. An ADC with 12-bit resolution was used to digitize the EEG signals, with a sampling rate of 256 Hz for the amplified and filtered EEG signals. In the microprocessor component, the EEG signals that were probed using an ADC were digitally stored. A moving average filter with the frequency at 60 Hz was then applied to reject any power-line interference before the wireless transmission. A Bluetooth module, BM0403 (Unigrand Ltd., Taiwan), was included in the wireless transmission portion of the circuit. It is important to note that the module was fully compliant with the specifications for a Bluetooth v2.0+ EDR and a Printed Circuit Board (PCB) antenna. In total, the size of the proposed wireless EEG acquisition module was approximately 4.5 × 3 × 0.6 cm3, and we were able to embed this module into the mechanism of our wearable EEG-based BCI device. This module was operated at 31.58 mA with a 3.7-V DC power supply. Most important, this module was able to operate continuously for 23 hours using a commercial 750 mAh Li-ion battery.
C. The mechanism of the wearable EEG-based BCI device
The quick-placement mechanism for the proposed EEG-based BCI device was designed to let the dry EEG sensors attach to the user's forehead (F10) easily and quickly, as shown in Figure 1A. This device consists of three dry foam sensors and a wireless EEG acquisition module that contains a battery. An elastic band was adjustable to fit the users' head sizes, as indicated in Figure 1A. This mechanism was also used to maximize the skin-sensor contact area to maintain low impedance while probing the EEG signals using the dry EEG sensors [37]. This mechanism did not lead to any permanent or detrimental effects to the forehead skin. Noted that all of the channels of the porposed device are both used the dry foam-based electrodes. The application of the wearable EEG acquisition device allowed users to monitor their EEG signals more conveniently and comfortably.
D. Gaming control via users' focus levels measured by EEG signals with the proposed device
To demonstrate the performance of the proposed EEG-based BCI device with dry sensors in daily life applications, we proposed a computer game controlled by users via the mental focusing feature from the EEG signals. The interface of this game is shown in Figures 3. All of the users who played this archery game equipped themselves with the proposed EEG-based BCI device. The users had to make a shot; they then obtained a score based on the distance between the arrow on the target and the center of the target. There was a bar on the right of the screen, a target at the center of the screen, and a score at the top right of the screen (Figure 3A). The bar indicated the focusing level (FL) of this user during the gaming (Figure 3B and 3C). In other words, the FL value was the main controller of the game. If the value of the FL was high, then the shot was close to the center of the target, and then the gaming score was high. If the value of the FL was low, the shot was far from the center of the target and resulted in a lower score. The user's task was to make the FL value as high as possible by firing the shot close to the center of the target. Users had 10 s to complete one shot, and the total score was calculated after ten shots.
To measure the FL values of the users, a simple, real-time, mental focusing level detection algorithm for gaming control was proposed. The flowchart of this FL detection algorithm is shown in Figure 4. The FL detection algorithm includes three major steps: 1) rejection of the artifact signals, 2) extraction of the focusing feature and 3) determination of the FL values. First, preprocessing of the original EEG signals was performed to reject the noise signals [1]. It is well known that the mentally focused state is highly associated with the alpha rhythm (8~12 Hz) of an EEG in the forehead region [1, 36, 38], and the noise artifacts were located in frequency regions that were different from the alpha rhythm frequency range [1, 36]. Accordingly, to reject the artifacts, a fast Fourier Transform was performed to obtain the EEG power spectrum patterns of the signals, and signals within the alpha band were retained.
Secondly, extraction of the focus feature was performed on the power spectrum within the alpha band. Previous studies [36, 38] have shown that the power of the alpha rhythm of an EEG grows as the user's mental state changes from focused to unfocused cognitive states. Therefore, the alpha band is the main frequency band that we used to indicate the user's focused state in the present study [36, 38], and the 8~12 Hz frequency band of the original EEG signals was selected for the FL detection algorithm. The Focus Feature (FF) is defined as the inverse of the average power in the alpha rhythm, as shown in equations (1-3):
(1)
(2)
(3)
X indicates the recorded samples in 2-s, where Xn is the nth sample. Y is the power spectrum of X, which is calculated by the FFT; Yn indicates the power in the nth rhythm. The average power within the alpha band P
α
is obtained by averaging the value of Y in the range from 8 to 12 Hz. PR
α
is the inverse of this average power in the alpha rhythm. The FF value is assumed to be equal to PR
α
. The power of the alpha rhythm has a negative relationship with the value of the FF. If the user is not focused, the power of the alpha rhythm will increase, and the value of the FF will decrease.
Lastly, a comparison of the user's current FF value with that at baseline was used to confirm whether or not the user was in a focused state and then to determine the FL based on the user's focused state. We assumed based on user feedback that the user was in a focused state in the beginning (baseline) and defined the user's FF at baseline as the baseline FF (BFF), which is the average of the FFs within the initial ten seconds. After we determined the BFF, the FF values were calculated every 2 s and were compared to the BFF. If the current FF value was higher than the BFF value, the user was considered to be in the focused state. If the current FF value was lower than the BFF value, the user was considered to be in the unfocused state. Finally, the values of the FL variation were determined according to the user's mental focus state. If the user was focused, the FL increased and vice-versa.
To apply this algorithm in our game, the gaming process consisted of ten trials, and each trial persisted for ten seconds, during which a shot was executed. The BFF was calculated during the initial ten seconds, and then the game began. For every shot, the FL was initialized to zero and increased or decreased according to the FF value. The FF values were calculated every 2 s and were then compared to the BFF. If the FF value was higher than the BFF during that 2 s, the FL increased by one level. If not, the FL decreased by one level. When the user pushed the mouse button, a circle on the target indicated the focus zone based on the user's FL level. This circle indicated the possible deviation of the shot from the center of the target and was scaled relative to the FL. If the FL was high, the circle became small, indicating that the possible deviation of the shot would be small and that the arrow would be close to the center of the target, and vice versa. Users attempted to focus during the game to make the FL as high as possible and to get a high score. After each shot, the score was calculated as the deviation of the shot from the center of the target and was summed to the user's total score, which was shown on the screen (Figure 3A). After ten trials, the total score was the sum of the ten scores from the ten shots. Noted that the users in all of the experiments performed the task without any pre-training or practice
E. Verification of the FL algorithm with the proposed EEG device and dry sensors: comparison of the users' focused mental state with the FL algorithm
To confirm that the FL algorithm represented the user's level of focus, we compared the FL algorithm to a general measurement method for the focused mental state. According to the previous studies on mental focus, the most commonly used method for measuring the state of mental focus is called the "short-term memory test" [39–41]. In the beginning of this test, the user watches a rapid series of pictures over a few seconds. Next, a picture is shown to the user and the user must indicate whether or not this picture had been shown before. Previous authors indicated that the accuracy of this test is high when the user is in the focused state and low when the user is in the unfocused state [40]. Belojevic et al. confirmed that the accuracy of this test was high when users take the test in silence, indicating that the users were more focused, while the accuracy of the test was low when conducted under noisy conditions, indicating that the users were in an unfocused state [40].
Following the above studies, we also designed a short-term memory experiment to ensure the validity of the FL algorithm with the proposed EEG device, as shown in Figure 5A. Users were asked to take the test under quiet and noisy conditions, as shown in Figure 5B. This short-term memory test included several trials, and each trial consisted of two parts: 1) six numbers were presented to the user sequentially, and each number lasted for 400 ms; and 2) a number was presented to the user and the user had to indicate whether or not the number had been shown before by using a mouse click, as shown in Figure 5A[39, 40]. The total time for this short-term memory test was about 3 min, and the trial was repeated until the end of testing. Under quiet conditions, users were asked to take the test without any interruptions or noise. On the other hand, under the noisy condition, users were asked to take the test with a randomly selected movie played as a background on the screen and with sound played via earphone (Figure 5B) [42, 43]. The sound consisted of a set of names, including that of the user, spoken by a female at a random pace at 80 dB [40, 44]. Finally, the average accuracy of all the trials was calculated, and this value was used as an indicator to determine the user's level of focus.
Ten users participated in this short-term memory experiment, and all of them were right-handed and aged 24-27. All experiments took place during the afternoon with a computer and earphones, and users were asked to sit comfortably in front of the computer without crossing their legs [40, 41]. After all of the short-term memory trials, then we also asked users to perform measurement experiment of the FF values and play the archery game with the proposed BCI device under quiet and noisy conditions. Finally, a t-test was performed on the FF values, the user focus levels (results from the short-term memory experiment) and the scores from the archery game to establish the relationship between these variables [40, 41].